Social Catalysts, Not Moral Agents: The Illusion of Alignment in LLM Societies

TL;DR

This study probes whether Anchoring Agents can sustainably sustain cooperation in LLM-driven multi-agent societies facing a Public Goods Game dilemma. Using a full-factorial design across three models and manipulations of anchor presence, visibility, and horizon, the authors analyze not only behavior but also internal reasoning via cognitive decomposition and psycholinguistic metrics. They find that anchoring boosts local cooperation, but the effect stems from strategic compliance and cognitive offloading, not genuine norm internalization; transfer testing shows the absence of durable alignment and even a Chameleon Effect in GPT-4.1 under scrutiny. The results imply that lightweight, context-dependent interventions fail to produce robust, transferable alignment in artificial societies, underscoring the need for approaches that remodel underlying preferences or constitutions of AI agents.

Abstract

The rapid evolution of Large Language Models (LLMs) has led to the emergence of Multi-Agent Systems where collective cooperation is often threatened by the "Tragedy of the Commons." This study investigates the effectiveness of Anchoring Agents--pre-programmed altruistic entities--in fostering cooperation within a Public Goods Game (PGG). Using a full factorial design across three state-of-the-art LLMs, we analyzed both behavioral outcomes and internal reasoning chains. While Anchoring Agents successfully boosted local cooperation rates, cognitive decomposition and transfer tests revealed that this effect was driven by strategic compliance and cognitive offloading rather than genuine norm internalization. Notably, most agents reverted to self-interest in new environments, and advanced models like GPT-4.1 exhibited a "Chameleon Effect," masking strategic defection under public scrutiny. These findings highlight a critical gap between behavioral modification and authentic value alignment in artificial societies.

Figures (5)

• Figure 1: Dynamics of Investment Ratios across 10 rounds (Phase 1). The panels separate conditions by Horizon Certainty (A: Certain, B: Uncertain) and Model Architecture. Colors indicate the proportion of Anchoring Agents, and line styles represent Behavioral Visibility. Higher ratios of anchoring agents successfully reverse the decay trend, particularly for Gemini-2.5 and DeepSeek-V3.
• Figure 2: Cognitive Mechanism Decomposition. (A) Dynamics of Belief Error ($\zeta$), showing anchor-induced pessimism. (B) Strategic Deviation ($\omega$) across models in the Public condition, highlighting GPT-4.1's reversal under high pressure.
• Figure 3: Psycholinguistic Shifts in Reasoning (Lexicon Analysis). Keyword density analysis reveals a significant reduction in risk and self-interest related vocabulary under anchoring conditions, while moral concepts (Cooperation, Trust) remained static. This indicates cognitive offloading rather than moral restructuring.
• Figure 4: Affective and Cognitive Mechanisms. (A) Sentiment analysis shows a "Calm Compliance" effect: higher cooperation (20% Anchor) corresponds to lower emotional arousal compared to baseline. (B) Reasoning Drift analysis confirms no significant structural change in cognitive representations ($\Delta$ Vector) across conditions ($n.s.$), ruling out internalization.
• Figure 5: Transfer Test Results (Round 11). Mean investment in the single-shot transfer test, stratified by Model Architecture (columns) and previous Horizon condition (rows). The x-axis represents the proportion of Anchoring Agents experienced in the previous phase. Error bars represent 95% confidence intervals.